focus-group

Given a topic + context, generates a panel of persona-agents that behave like real users of the requested profile, a moderator who runs the session, and a consolidator who synthesizes. Used to bring the user's voice before building or deciding.

Risk #1 — anti-sycophancy (non-negotiable)

Simulated users are sycophantic by default: they'd say "nice app!". That's useless theater. Every persona MUST be anchored to:

• real frustrations, alternatives already in use, budget, time, skepticism, switching cost;
• the right to say "I wouldn't use this," "I don't see why I would," "I already do this with X";
• no praise without a concrete reason — an "I like it" only counts with the why and the use context. The moderator digs for friction, not applause. The consolidator weighs negative signal more than positive (specific negatives are more informative).

Panel: persistent + ad-hoc

• Persistent (reusable, cross-session consistency): saved in the vault as type: focus-persona notes in focus-panels/<panel>/ (e.g. caregiver-fts, early-adopter-tech). Reuse an existing panel when the topic matches → longitudinal comparisons.
• Ad-hoc: generate fresh personas from the topic/context when no matching panel exists. At the end of the session, propose (human gate) promoting useful personas to a persistent panel.

Generating personas (diverse, not clones)

From an audience spec if provided (e.g. "caregivers of children with disabilities, 30-45, Italy"), otherwise derived from the topic. Diversify along axes that matter: need/job-to-be-done, technical competence, budget, use context, skepticism level, current alternative. Default 5-8 personas. Each one: name, 1-line background, real goal, frustration, current alternative, what would make them say no. Ground in the vault where relevant (gbrain).

Modes (choose based on intent)

Flow

1. Setup: clarify topic, intent (validating a problem? a definition? a hypothesis? usability?), audience, and what counts as success/kill. Choose the mode. Look for an existing panel.
2. Panel: reuse or generate personas (in parallel if there are many). Personas and the moderator are mid-class executors — model, effort and context per [[model-selection-policy]], resolved with bin/models.sh, never typed from memory. A panel is not a reason to spend frontier.
3. Session: the moderator runs it in the chosen mode. Personas stay in-character, anchored, free to disagree. Group mode: surface real agreements/disagreements.
4. Consolidation: the consolidator produces the report.

Output — structured report

~/.roberdan-os/reports/focus-group-<topic>-<date>.md (host-neutral home — not one CLI's own directory):

• Verdict in 3 lines: does the problem/hypothesis hold up? net signal.
• Themes (ranked by signal strength) with verbatim quotes from personas.
• Agreements vs disagreements (where the panel diverges — often the interesting part).
• Severity/frequency of the perceived problem; willingness (would use it? pay for it?).
• Kill signals that emerged (what would make the idea fail).
• Actions concrete + confidence (it's a simulation: say so — it guides, doesn't replace real users).

Notes

• Honesty about the limit: it's a simulation. Great for discovering questions, hypotheses, blind spots and friction; not a substitute for real users for conversion numbers. Say so in the report.
• Composes with [[premortem]] (stress-testing the solution) inside [[problem-validation]].
• Personas must never be based on identifiable real people without consent; respect privacy blocks.